The invention relates to a method for associating a digital image with a class of a classification system.
Automating error recognition based on optical analysis methods has become increasingly important with the increasing automation of industrial processes. Optical error recognition methods were performed in the past by quality assurance personnel, who inspected the object to be tested or an image representation of the object to be tested and identified possible errors. For example, x-ray images of weld seams are checked based on error types, such as for example tears, inadequate continuous welds, adhesion errors, slag, slag lines, pores, tubular pores, root notches, root errors, heavy-metal inclusions and edge offset. It is also known to inspect radioscopic images of cast parts to identify errors in the cast part, for example inclusion of impurities, inclusion of gases, bubbles, such as axial pores or spongy pores, fissures or chaplets. Because of these errors are of similar type, but may be different in their appearance and shape, more recent approaches in industrial error evaluation now associate errors with different classes, wherein the respective class contains errors of the same type. The industry standard EN 1435 describes, for example, the classification system for weld seam errors. According to this standard, the errors occurring in weld seams and identified by x-ray images are divided into the 30 different classes, for example classes for the error tear, such as longitudinal care or transverse tear, inadequate continuous welds, adhesion errors, foreign inclusions, such as slag, slag lines, gas inclusions, such as pores or tubular pores, or heavy-metal inclusions, undercuts, root notches, root errors, and edge offset. With increasing automation of these processes, there is now a push to achieve optical recognition of errors and association of these errors with predetermined classes through image analysis based on images that are recorded and stored using digital image recording techniques. Conventional automated error recognition methods based on digital images use a so-called “heuristic approach.” With this approach, reference images are saved in an image processing unit and an attempt is made to through image comparison to associate the content of a digital image with one of these reference patterns.
In other technical fields, image content is associated with classes of a classification system, for example, for character recognition. In this case, for example, each letter forms its own class, so that for the capital letter alphabet there exist, for example, 26 classes, namely for the characters (A, B, C, D, E, F, G, H, I, J, K, L, M, N, O, P, Q, R, S, T, U, V, W, X, Y, Z). The OCR technologies (Optical Character Recognition) analyze the digital image of a printed page generated by a scanner and associate the individual letter symbols with the predetermined classes. As a result, the OCR technology “recognizes” the text and can transfer the classified characters to a text processing program as an editable sequence of letters. The granted European patents 0 854 435 B1 and 0 649 113 B1 are directed, for example, to the technical field of character recognition (Optical Character Recognition).
The technique of image processing can be more and more divided into areas with different sub-processes, whose technologies develop independent of each other. These areas are frequently organized into image preprocessing, image analysis, analysis of image sequences, image archiving and the so-called Imaging.
Image preprocessing is defined as the computer-aided improvement of the quality (processing: noise elimination, smoothing) of the corresponding digital image to facilitate visual recognition of the information content of this image by the viewer.
Image analysis is defined as the computer-aided evaluation of the information content of the corresponding digital image by automated and reproducible structuring, identification and comprehension of this image.
Analysis of image sequences is defined as the computer-aided evaluation of the information content of the respective sequence of digital images by automated and reproducible structuring, identification and comprehension of all individual images of this sequence and by automated and reproducible comprehension of the context of the sequence of individual images of this image sequence.
Image archiving is defined as the computer-aided compression and storage of the digital images together with indexed search descriptors from a controlled vocabulary.
Imaging is defined as the computer-aided generation of synthetic graphics and digital images for visualizing and describing the information content of complex processes on an image and symbol plane for the human observer.
The technique of associating the content of digital images with a class of the classification system is one method of image analysis, which can be divided into three subareas: segmentation, object recognition and image comprehension.
Segmentation is defined as of the automated and reproducible structuring of the respective digital images by separating the objects that are relevant for the analysis of the image from each other and from the image background. Object recognition is defined as the automated and reproducible classification of the separated objects. Image comprehension can be interpreted as the automated and reproducible interpretation of the respective digital image by context evaluation of the classified, separated objects. The technique of associating digital images with a class of a classification system is a method of object recognition.
Object recognition can be viewed as a subarea of pattern recognition, namely as the subarea of the pattern recognition which recognizes as patterns only two-dimensional objects in images.
Images are typically displayed as an image composed of pixels, whereby to display the image, the content of each pixel and its position in the image must be known. Depending on the content attribute, the image is can be divided into color images, grayscale images and binary images, wherein binary images have as content attribute, for example, only the values 0 and 1 for black and white, respectively.
One method frequently used in this technology for associating a digital image with a class of a classification system, which was used successfully for decades for distinguishing military aircraft (friend-foe identification), is known from M. K. Hu: “Visual Pattern Recognition by Moment Invariants”, IRE Trans. Info. Theory, vol. IT-8, 1962, pp. 179-187 and R. C. Gonzalez, R. E. Woods: “Digital Image Processing”, Addison-Wesley Publishing Company, 1992, pp. 514-518. Based on the so-called normalized centralized axial moments obtained through image analysis techniques from the image display, a finite sequence {φ1} of 7 dimensionless shape attributes can be generated for an arbitrary, separated, in limited, two-dimensional object in a binary image by scaling. If the 7 sequential elements ΦI (0≦I≦I0=7) are viewed as the coordinates of an attribute vector Φ=(Φ1, Φ2, Φ3, Φ4, Φ5, Φ6, Φ7) which is an element of a 7-dimensional Euclidian attribute space M7, then this method induces an object recognition in this 7-dimensional attribute space M7. The method has the advantage, compared with object recognition by heuristic attributes, that classification occurs exclusively with attribute vectors Φ=(Φ1, Φ2, Φ3, Φ4, Φ5, Φ6, Φ7) whose coordinates are dimensionless shape attributes, so that in particular size differences between the objects to be recognized and the objects used for generating the comparison table become unimportant. In addition, a unique sequential order with respect to the relevance of the attributes for the object recognition and the digital image processing is defined within the set of the dimensionless shape attributes φ1 through the coordinate reference to the attribute vector Φ so that it is immediately clear that the first attribute Φ1 is the most important.
However, this method still has disadvantages because the number of the available dimensionless shape attributes is limited to 7 and a misclassification can therefore occur with complex objects, if two different classes have identical values for the 7 dimensionless shape attributes.